CN117423396A - Crystal structure generation method and device based on diffusion model - Google Patents
Crystal structure generation method and device based on diffusion model Download PDFInfo
- Publication number
- CN117423396A CN117423396A CN202311737478.8A CN202311737478A CN117423396A CN 117423396 A CN117423396 A CN 117423396A CN 202311737478 A CN202311737478 A CN 202311737478A CN 117423396 A CN117423396 A CN 117423396A
- Authority
- CN
- China
- Prior art keywords
- crystal structure
- diffusion model
- noise
- data set
- structure generation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 239000013078 crystal Substances 0.000 title claims abstract description 326
- 238000009792 diffusion process Methods 0.000 title claims abstract description 162
- 238000002910 structure generation Methods 0.000 title claims abstract description 118
- 238000000034 method Methods 0.000 title claims abstract description 113
- 238000004925 denaturation Methods 0.000 claims abstract description 8
- 230000036425 denaturation Effects 0.000 claims abstract description 8
- 230000008569 process Effects 0.000 claims description 65
- 238000012549 training Methods 0.000 claims description 60
- 238000012545 processing Methods 0.000 claims description 25
- 238000012360 testing method Methods 0.000 claims description 23
- 238000006243 chemical reaction Methods 0.000 claims description 22
- 238000013528 artificial neural network Methods 0.000 claims description 21
- 230000006870 function Effects 0.000 claims description 21
- 238000012795 verification Methods 0.000 claims description 15
- 239000002131 composite material Substances 0.000 claims description 12
- 239000011159 matrix material Substances 0.000 claims description 12
- 238000000547 structure data Methods 0.000 claims description 8
- 239000000126 substance Substances 0.000 claims description 8
- 230000005540 biological transmission Effects 0.000 claims description 6
- 238000012512 characterization method Methods 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 6
- 239000013589 supplement Substances 0.000 claims description 3
- 230000001502 supplementing effect Effects 0.000 claims description 2
- 238000005457 optimization Methods 0.000 abstract description 5
- 230000005484 gravity Effects 0.000 abstract description 4
- 229920002994 synthetic fiber Polymers 0.000 abstract description 2
- 125000004429 atom Chemical group 0.000 description 25
- 230000008901 benefit Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 239000000463 material Substances 0.000 description 4
- 102100026038 Lens fiber membrane intrinsic protein Human genes 0.000 description 3
- 101710115990 Lens fiber membrane intrinsic protein Proteins 0.000 description 3
- 230000015572 biosynthetic process Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000003786 synthesis reaction Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000004590 computer program Methods 0.000 description 2
- 238000013434 data augmentation Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 239000007809 chemical reaction catalyst Substances 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- HNPSIPDUKPIQMN-UHFFFAOYSA-N dioxosilane;oxo(oxoalumanyloxy)alumane Chemical group O=[Si]=O.O=[Al]O[Al]=O HNPSIPDUKPIQMN-UHFFFAOYSA-N 0.000 description 1
- 238000007876 drug discovery Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 125000004430 oxygen atom Chemical group O* 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C10/00—Computational theoretical chemistry, i.e. ICT specially adapted for theoretical aspects of quantum chemistry, molecular mechanics, molecular dynamics or the like
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/70—Machine learning, data mining or chemometrics
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C60/00—Computational materials science, i.e. ICT specially adapted for investigating the physical or chemical properties of materials or phenomena associated with their design, synthesis, processing, characterisation or utilisation
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Chemical & Material Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Crystallography & Structural Chemistry (AREA)
- Databases & Information Systems (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Medical Informatics (AREA)
- Analysing Materials By The Use Of Radiation (AREA)
Abstract
The crystal structure generation method and device based on the diffusion model not only can comprehensively represent the E (3) and other denaturation of the crystal by utilizing the relative distance between atoms, but also does not change the data set; the linear interpolation is introduced to improve the diversity and novelty of crystal generation, the linear combination mode is used to respectively add Gaussian noise to two crystal structures according to specific gravity to fuse the two crystal structures to obtain mixed Gaussian noise, then the noise is gradually removed, the two crystal structures can be fused smoothly and controllably, tens of thousands of novel crystal structures can be generated theoretically, the cost of artificial synthetic materials is greatly saved, and the conception of a synthetic mode is omitted; the key predictor is used as a guide to improve the generation precision of the atomic coordinates, and based on the strong dependence of the key on the atomic coordinates, the slight atomic disturbance can influence the type of the key and whether the key is formed, so that the cross entropy loss function is used and the gradient is used as an optimization direction to correct the generation of the atomic coordinates; comprehensively improving the expression of the diffusion model in the generation of inorganic crystal structures.
Description
Technical Field
The invention relates to a crystal structure generation method and device based on a diffusion model, and belongs to the technical field of inorganic crystal structure generation.
Background
At present, diffusion models have been applied to various image generation scenes, and have been gradually expanded to natural science fields such as drug discovery, material design, medical image reconstruction, and the like. In molecular generation, especially organic molecular generation, depth generation methods based on diffusion models are also attracting attention of researchers.
The inorganic crystal material has very rich functions, can be used for preparing various devices such as lasers, LEDs, solar cells, sensors and the like, and can also be used as a gas-solid reaction catalyst. Diffusion models for inorganic crystal structure generation are in the initial stage of research, with naive properties in terms of handling crystal structure symmetry and the manner in which bonds are generated. The traditional technology adopts a data augmentation mode to represent the E (3) invariance of a crystal structure, has poor robustness and generalization capability, and amplifies a training data set by a plurality of times, thereby increasing the consumption of computing resources, and is not optimistic in both training precision and convergence speed.
Disclosure of Invention
Therefore, the invention provides a crystal structure generation method and device based on a diffusion model, which solve the problems of poor robustness and generalization capability, high calculation resource consumption and poor generation effect of the traditional technology.
In order to achieve the above object, the present invention provides the following technical solutions: the method for generating the crystal structure based on the diffusion model comprises the following steps:
acquiring an open source crystal structure data set as a crystal structure generation data set, wherein the crystal structure generation data set is composed of CIF type files representing unit cell parameters and atomic three-dimensional coordinates of each crystal structure;
converting the CIF type file batch into SDF type files required by training by using a chemical information file type conversion package, wherein the SDF type files comprise the number of unit cell atoms, an atom bonding adjacent matrix and three-dimensional coordinates of each atom; dividing the crystal structure generated data set after format conversion into a training set, a verification set and a test set;
constructing a crystal structure generation diffusion model, wherein the crystal structure generation diffusion model is provided with a forward noise adding process and a reverse noise removing process, the forward noise adding process uses predefined hyper-parameters to represent the mean value and variance of conditional probability distribution, and the reverse noise removing process restores data by predicting noise added by the forward noise adding process;
constructing a key predictor through a graph neural network, wherein the input value of the key predictor is the atomic coordinate and the atomic type in the data set generated by the crystal structure, and the output value of the key predictor is the key type;
Training the crystal structure generation diffusion model by using a training set, evaluating the prediction capability of the training crystal structure generation diffusion model in an unknown data set by using a verification set, and testing the trained crystal structure generation diffusion model by using a test set to obtain a final crystal structure generation diffusion model;
generating a diffusion model by using a given Gaussian noise, and generating a crystal structure by the constructed bond predictor and the final crystal structure;
and performing linear interpolation on the novel crystal structure generated by the diffusion model generated by the trained crystal structure, adding noise to different time stamps from two different crystal structures, and setting interpolation factors to control the occupation proportion of each crystal structure in the composite crystal.
As a preferred scheme of the crystal structure generation method based on the diffusion model, performing diffusion treatment on the crystal structure generation data set subjected to format conversion through the crystal structure generation diffusion model, wherein the crystal structure generation data set subjected to diffusion treatment approximately accords with standard Gaussian distribution;
in the reverse denoising process, a neural network is constructed to learn and predict added noise, the relative distance between atoms is used for carrying out denaturation such as E (3) of a transmission characterization crystal, and the L2 norm between the added noise and the predicted noise is used as a loss function.
As a preferable scheme of the crystal structure generation method based on the diffusion model, the output value key type of the key predictor is judged to be a classification task, cross entropy is used as a loss function, the crystal structure is guided through gradient values to generate learning condition probability distribution of the diffusion model, and generated crystal atomic coordinates are corrected.
As a preferred embodiment of the method for generating a crystal structure based on a diffusion model, a given gaussian noise is used to generate a crystal structure by generating a diffusion model from a final crystal structure:
randomly collecting noise values from multidimensional standard Gaussian noise distribution as a starting value to realize a reverse denoising process, wherein the Gaussian noise dimension of the reverse denoising comprises atomic coordinates contained in a unit cell and one-hot codes of all atomic types;
and continuously denoising the input value by the crystal structure generation diffusion model according to the learned conditional distribution, and finally obtaining a true value conforming to the distribution of the original data set.
As a preferred scheme of the crystal structure generation method based on the diffusion model, the crystal structure generation diffusion model is added through linear interpolation to generate a crystal structure type, two different crystal structures are respectively noisy to time stamps T1 and T2, and t1+t2=t, wherein T is the total diffusion time step.
As a preferred scheme of the crystal structure generation method based on the diffusion model, the generated crystal structure is used as a training set to supplement, and the diffusion model generated by the crystal structure is trained again, so that the generated crystal structure accords with the distribution of the original data set.
The invention also provides a crystal structure generating device based on the diffusion model, which comprises:
the system comprises a data set acquisition module, a crystal structure generation module and a crystal structure generation module, wherein the data set acquisition module is used for acquiring an open source crystal structure data set as a crystal structure generation data set, and the crystal structure generation data set is composed of cell parameters representing each crystal structure and CIF type files of atomic three-dimensional coordinates;
the format conversion module is used for converting the CIF type file batch into SDF type files required by training by using a chemical information file type conversion package, wherein the SDF type files comprise the number of unit cell atoms, an atom bonding adjacent matrix and three-dimensional coordinates of each atom;
the data set dividing module is used for dividing the crystal structure generated data set after format conversion into a training set, a verification set and a test set;
the diffusion model processing module is used for constructing a crystal structure generation diffusion model, the crystal structure generation diffusion model is provided with a forward noise adding process and a reverse noise removing process, the forward noise adding process uses a predefined hyper-parameter to represent the mean value and variance of conditional probability distribution, and the reverse noise removing process restores data by predicting noise added by the forward noise adding process;
The key predictor construction module is used for constructing a key predictor through the graph neural network, wherein the input value of the key predictor is the atomic coordinate and the atomic type in the data set generated for the crystal structure, and the output value of the key predictor is the key type;
the model training module is used for training the crystal structure generation diffusion model by using the training set, evaluating the prediction capability of the training crystal structure generation diffusion model in an unknown data set by using the verification set, and testing the trained crystal structure generation diffusion model by using the test set to obtain a final crystal structure generation diffusion model;
and the crystal structure generation module is used for generating a crystal structure by utilizing given Gaussian noise through the constructed bond predictor and the final crystal structure generation diffusion model.
And the linear interpolation processing module is used for carrying out linear interpolation on the novel crystal structure generated by the diffusion model generated by the trained crystal structure, adding noise to different time stamps from two different crystal structures, and setting interpolation factors so as to control the occupation proportion of each crystal structure in the composite crystal.
As a preferred solution of the crystal structure generating device based on the diffusion model, in the diffusion model processing module, the diffusion processing is performed on the crystal structure generating data set after format conversion through the crystal structure generating diffusion model, and the crystal structure generating data set after diffusion processing approximately accords with standard gaussian distribution;
In the reverse denoising process of the diffusion model processing module, a neural network is constructed to learn and predict added noise, the relative distance between atoms is used for carrying out denaturation such as E (3) of a transmission characterization crystal, and the L2 norm between the added noise and the predicted noise is used as a loss function.
As a preferable mode of the crystal structure generating device based on the diffusion model, in the key predictor construction module, the output value key type of the key predictor is judged to be a classification task, cross entropy is used as a loss function, the crystal structure is guided through gradient values to generate learning condition probability distribution of the diffusion model, and the generated crystal atomic coordinates are corrected.
As a preferred scheme of the crystal structure generating device based on the diffusion model, in the crystal structure generating module, noise values are randomly collected from multidimensional standard Gaussian noise distribution as initial values to realize a reverse denoising process, and the Gaussian noise dimension of the reverse denoising comprises atomic coordinates contained in a unit cell and one-hot codes of all atomic types;
in the crystal structure generation module, the diffusion module is generated through the crystal structure to continuously denoise the input value according to the learned condition distribution, and finally, the true value conforming to the distribution of the original data set is obtained;
In the linear interpolation processing module, the crystal structure generation diffusion model is added through linear interpolation to generate crystal structure types, two different crystal structures are respectively noisy to time stamps T1 and T2, and t1+t2=t, wherein T is the total diffusion time step;
the model retraining module is used for supplementing the generated crystal structure into a training set, and retraining the diffusion model generated by the crystal structure so that the generated crystal structure accords with the distribution of the original data set.
The invention has the following advantages: the relative distance between atoms is utilized from the essence, so that not only can the E (3) and other denaturation of the crystal be comprehensively represented, but also the data set is not changed; the diversity and novelty of crystal generation are improved by introducing a linear interpolation mode, the two crystal structures are fused by adding Gaussian noise according to specific gravity respectively in a linear combination mode to obtain mixed Gaussian noise, then the noise is gradually removed, the two crystal structures can be fused smoothly and controllably, tens of thousands of novel crystal structures can be generated theoretically, the cost of artificial synthetic materials is greatly saved, and the conception of a synthetic mode is omitted; the key predictor is used as a guide to improve the generation precision of the atomic coordinates, and based on the strong dependence of the key on the atomic coordinates, the slight atomic disturbance can influence the type of the key and whether the key is formed, so that the cross entropy loss function is used and the gradient is used as an optimization direction to correct the generation of the atomic coordinates; comprehensively improving the expression of the diffusion model in the generation of inorganic crystal structures.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. It will be apparent to those skilled in the art from this disclosure that the drawings described below are merely exemplary and that other embodiments may be derived from the drawings provided without undue effort.
FIG. 1 is a schematic flow chart of a method for generating a crystal structure based on a diffusion model according to an embodiment of the present invention;
fig. 2 is a diagram of a crystal structure generating device according to an embodiment of the present invention.
Detailed Description
Other advantages and advantages of the present invention will become apparent to those skilled in the art from the following detailed description, which, by way of illustration, is to be read in connection with certain specific embodiments, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1
Referring to fig. 1, an embodiment of the present invention provides a crystal structure generating method based on a diffusion model, including the steps of:
s1, acquiring an open source crystal structure data set as a crystal structure generation data set, wherein the crystal structure generation data set is composed of CIF type files representing unit cell parameters and atomic three-dimensional coordinates of each crystal structure;
s2, converting the CIF type file batch into SDF type files required by training by using a chemical information file type conversion package, wherein the SDF type files comprise the atomic number of unit cells, an atomic bonding adjacent matrix and three-dimensional coordinates of each atom; dividing the crystal structure generated data set after format conversion into a training set, a verification set and a test set;
s3, constructing a crystal structure generation diffusion model, wherein the crystal structure generation diffusion model is provided with a forward noise adding process and a reverse noise removing process, the forward noise adding process uses a predefined hyper-parameter to represent a conditional probability distribution mean value and a variance, and the reverse noise removing process restores data by predicting noise added by the forward noise adding process;
s4, constructing a key predictor through a graph neural network, wherein the input value of the key predictor is the atomic coordinate and the atomic type in the data set generated by the crystal structure, and the output value of the key predictor is the key type;
S5, training the crystal structure generation diffusion model by using a training set, evaluating the prediction capability of the training crystal structure generation diffusion model in an unknown data set by using a verification set, and testing the trained crystal structure generation diffusion model by using a test set to obtain a final crystal structure generation diffusion model;
s6, generating a diffusion model to generate a crystal structure through the constructed bond predictor and the final crystal structure by using given Gaussian noise;
s7, performing linear interpolation on the novel crystal structure generated by the trained crystal structure generated diffusion model, adding noise to different time stamps for two different crystal structures, and setting interpolation factors to control the occupation proportion of each crystal structure in the composite crystal.
In this embodiment, in step S1, the open-source zeolite structure data set MP20/PCOD is used as the crystal structure generation data set, which is composed of CIF type files containing information such as the number of Si and O atoms, three-dimensional coordinates, unit cell parameters, and the like. Training was performed using an open source crystal structure MP20 dataset containing 89 elements and the file format was CIF type.
In this embodiment, in step S2, the original crystal structure generation data set file format is converted to obtain a data format that can be read by the diffusion model. First, the zeo++ software is used to reduce the symmetry of the crystal to P1 for CIF type file processing, because the various symmetries of the crystal structure, the complete unit cell can represent its basic structure with minimal asymmetric units, which results in atoms that are not all connected, while the diffusion model represents the crystal structure with full connection diagram, thus the model training is not favored and the symmetry operation needs to be eliminated. The CIF type file format is converted into an SDF type file format representing the atomic number of each crystal structure, three-dimensional coordinates, and a key adjacency matrix. Crystal structure generation dataset according to scale 8:1:1 is divided into a training set, a verification set and a test set.
In this embodiment, in step S3, the crystal structure generating diffusion model is divided into a forward noise adding process and a reverse noise removing process, the forward noise adding process uses a predefined hyper-parameter to represent the mean and variance of the conditional probability distribution, is a mathematical fixed expression and does not need training, in this embodiment, the diffusion time setting step t=1000,= 10-4，/>=0.02 and->The linear reduction, the final post-diffusion dataset approximately conforms to a standard gaussian distribution.
The inverse denoising process restores data by predicting noise added by the previous forward denoising process, and because given one denoising data is unknown to the added noise, a neural network needs to be constructed for learning so as to predict the added noise, and a graph neural network is usedThe relative distance between the atoms of the complex pair is transmitted to characterize the E (3) and other denaturation of the crystal, and the added noise is generatedAnd predict noise->The L2 norm in between as a loss function.
Specifically, in step S4, the data of the crystal structure generation diffusion model are characterized as follows:
for atomic number ofNThe crystal M of (2) is represented as:
；
；
；
；
in the method, in the process of the invention,respectively representing the atom type, the atom coordinates and the adjacency matrix value, < >>Representing the type of element to which atom i belongs, +.>Representing the three-dimensional coordinates of atom i >Indicating whether atom i, j is bonded, the bonding value is 1, otherwise 0.
The forward noise-adding process is as follows:
；
in the noise parameter，/>Is an identity matrix, from the original structure->One step noise up to time stamp t (+)>) Can be expressed as:
；
in the method, in the process of the invention,representation->。
When (when)When (I)>Becomes a standard normal distribution.
Wherein, reverse denoising process uses neural network to train and satisfies:
；
in the method, in the process of the invention,representing the predicted mean>Representing time tCrystals at the time of stamping.
The final loss function is:
；
in the method, in the process of the invention,representing coincidence probability distribution->Is>Representing variance->Indicating added noise->Representing the predicted noise.
In this embodiment, in step S4, the output value key type of the key predictor is determined as a classification task, the cross entropy is used as a loss function, the gradient value is used to guide the crystal structure to generate a learning condition probability distribution of the diffusion model, and the generated crystal atomic coordinates are corrected.
Since the crystal structure has transformations (such as translation), symmetry, and rotational invariance in 3D space, it is also called E (3) and so on. The traditional characterization method is to amplify the data set, such as symmetrical operation, rotation operation and the like, but the method is poor in robustness and generalization capability, and the number of training sets is multiplied.
In the embodiment, the isomorphous graph neural network is introduced, so that not only can the E (3) isomorphous property of the crystal structure be well represented, but also the original data set can be kept unchanged. In addition, the information transmission of chemical bonds is considered, and whether bonds are formed is determined by adopting the previous logic judgment instead of the mechanical judgment, so that the whole space information is comprehensively utilized.
The graph neural network is characterized in the following manner:
；
；
；
；
；
；
in the method, in the process of the invention,respectively representing an atomic type single-heat code, a key feature (represented as an adjacent matrix value when no feature exists) and an atomic three-dimensional coordinate, +.>Representing the fused key features, linear () represents the Linear transformation of the corresponding item, ++>Are each a neural network composed of Multiple Layer Perceptrons (MLPs). The core part of the network is to use the relative distance between two atoms as input and to take into account all the bond information that two adjacent atoms are each connected, so that the crystal structure guarantees the relative position between the atoms unchanged regardless of the transformation, and to take into account the spatial information of the whole crystal. Finally, the atom type and whether a bond is formed is obtained by a softmax function while considering the symmetry of the bond. Thus constructing another graph neural network to construct a key predictor, wherein the input value is an atomic coordinate and an atomic type, and the output value is a key type If the key types of the data set are multiple, the data set is classified into multiple tasks, and if the key types are only one, the data set is classified into two tasks, cross entropy is used as a loss function, and gradient values of the cross entropy are used for guiding a diffusion model to learn conditional probability distribution so as to correct the generated crystal atomic coordinates, so that the effectiveness of generating a crystal structure is improved.
In this embodiment, in step S5, the training set is used to train the diffusion model generated by the crystal structure, and the training amounts of the MP20 data set and the PCOD data set are 35000 cases and 100000 cases, respectively. Wherein the diffusion noise precision is set to be 10 -5 10 -5 Noise superparameter is reduced linearly, training batch and epoch are 256 and 3000 respectively, and learning rate is 10 -4 The number of epochs tested was 20. And in the training process, the model generalization capability is tested through the cross validation model training effect and the test set respectively every 20 epochs.
In this embodiment, after the crystal structure generation diffusion model has the capability of generating a sample, the final purpose is to make the crystal structure generation diffusion model generate a conditional expression, just as a given text guide model generates an image conforming to semantics, and the accuracy of generating a crystal structure key by the crystal structure generation diffusion model is improved by a key predictor method.
Specifically, the strong association of bond length and whether bond is formed can guide the generation of atomic coordinates to play a role in correction, thereby training another graph neural networkPredicting all keys using atomic coordinates and type as input, assuming pairs of keysPredicted bond pair value is +.>And define a function +.>The confidence level is quantified. Thus all bonds in the unit cellThe confidence is:
；
if the generated atomic position is not accurate, the key predictor confidence is low, the gradient thereof can be used to establish a direction-optimizing objective function and set the super parameter s, at which time:
；
in the method, in the process of the invention,indicating confidence of all keys of the unit cell, +.>Representing the confidence level of the key ij,a predicted value representing the kth key type, a predicted value representing a key or a non-key when only the adjacency matrix,indicate a given +.>,/>Conditional probability distribution of->Representing the predicted mean.
In this embodiment, in step S6, in the process of generating a crystal structure by generating a diffusion model through a final crystal structure using a given gaussian noise:
randomly collecting noise values from multidimensional standard Gaussian noise distribution as a starting value to realize a reverse denoising process, wherein the Gaussian noise dimension of the reverse denoising comprises atomic coordinates contained in a unit cell and one-hot codes of all atomic types;
And continuously denoising the input value by the crystal structure generation diffusion module according to the learned conditional distribution, and finally obtaining a true value conforming to the distribution of the original data set.
Specifically, during the generation of the crystal structure, noise values are randomly collected from the multidimensional standard Gaussian noise distributionAs a starting value, the inverse denoising process is realized, and Gaussian noise dimensions comprise atomic coordinates contained in a unit cell, one-hot codes of all atomic types and the like, so that sample distribution of different sizes is consistent with original data set distribution when a sample is generated, and the generated crystal is ensured to be more approximate to the original data set distribution. From noise value->By distribution->Denoising in the T step to obtain a true value->。
The generated crystal structure part appears in the original data set, because when the training examples of the data set are too many, the noise of random sampling is inevitably overlapped with the standard Gaussian noise obtained by adding noise, and the crystal structure which does not appear in the data set is a novel crystal, thereby replacing the mechanical artificial synthesis attempt. In order to improve the novelty and diversity of crystal generation, a plurality of Gaussian noises are sampled simultaneously to obtain mixed Gaussian noises, and the mixed Gaussian noises are step by step denoised according to the method to generate the composite crystal. Furthermore, the method can be used for generating infinite crystal structures continuously, the generated crystal structures can be expanded to original data sets, the generated crystal structures are more accurate in an iterative mode, and the method can be applied to any crystal structure data set.
In this embodiment, in step S7, the crystal structure generation diffusion model is added by linear interpolation to generate a crystal structure type, and two different crystal structures are respectively noisy to time stamps T1, T2, where t1+t2=t, and T is the total diffusion time step.
Wherein, two different crystals increase noise from t=0 time stamp (original state) to t=t/2 time stamp, then add hidden variable space data to obtain mixed Gaussian noise, finally de-noise step by step from T times to t=0, and at the same time can set interpolation factor for each crystal structureThe specific gravity of each crystal structure is controlled so as to have flexibility without losing generality. Different interpolation factors represent smooth transitions between two different crystals, providing a new idea for designing new materials.
Specifically, the diversity and novelty of crystal structure generation are improved by a linear interpolation technology, and the composite material accords with the synthesis characteristic of the composite material to improve the superiority of crystals. Two different crystal structures are noisy to different time stamps T1, T2, and t1+t2=t (1000 steps), while the interpolation factor is set，/>And->The occupation proportion of each crystal structure in the composite crystal is controlled, so that mixed Gaussian noise is obtained, the model is trained according to the same training method to obtain the composite crystal structure, the method can be expanded to a mode of fusing a plurality of crystal structures, and a plurality of time stamps and interpolation factors are set.
In summary, in the embodiment of the present invention, an open source crystal structure dataset is obtained as a crystal structure generation dataset, where the crystal structure generation dataset is composed of CIF type files representing unit cell parameters and atomic three-dimensional coordinates of each crystal structure; converting the CIF type file batch into SDF type files required by training by using a chemical information file type conversion package, wherein the SDF type files comprise the number of unit cell atoms, an atom bonding adjacent matrix and three-dimensional coordinates of each atom; dividing the crystal structure generated data set after format conversion into a training set, a verification set and a test set; constructing a crystal structure generation diffusion model, wherein the crystal structure generation diffusion model is provided with a forward noise adding process and a reverse noise removing process, the forward noise adding process uses predefined hyper-parameters to represent the mean value and variance of conditional probability distribution, and the reverse noise removing process restores data by predicting noise added by the forward noise adding process; constructing a key predictor through a graph neural network, wherein the input value of the key predictor is the atomic coordinate and the atomic type in the data set generated by the crystal structure, and the output value of the key predictor is the key type; training the crystal structure generation diffusion model by using a training set, evaluating the prediction capability of the training crystal structure generation diffusion model in an unknown data set by using a verification set, and testing the trained crystal structure generation diffusion model by using a test set to obtain a final crystal structure generation diffusion model; and generating a diffusion model by using the given Gaussian noise through the constructed bond predictor and the final crystal structure to generate a crystal structure. And performing linear interpolation on the novel crystal structure generated by the diffusion model generated by the trained crystal structure, adding noise to different time stamps from two different crystal structures, and setting interpolation factors to control the occupation proportion of each crystal structure in the composite crystal. The traditional technology adopts a data augmentation mode to represent the E (3) invariance of a crystal structure, and the method is poor in robustness and generalization capability, and increases the calculation resource consumption by amplifying a training data set by a plurality of times, so that the method is not optimistic in both training precision and convergence speed. The invention adopts the method of the isomorphous graph neural network, from the essence, the relative distance between atoms is utilized, so that not only the E (3) isomorphous property of the crystal can be comprehensively represented, but also the data set is not changed; the invention utilizes the crystal structure to generate the diffusion model to represent the hidden variable space, the invention introduces a linear interpolation mode to improve the diversity and novelty of crystal generation, adopts a linear combination mode to respectively add Gaussian noise into two crystal structures according to specific gravity to fuse so as to obtain mixed Gaussian noise, then gradually removes the noise, and can smoothly and controllably fuse the two crystal structures. The method can theoretically generate tens of thousands of novel crystal structures, greatly saves the cost of artificially synthesized materials and omits the conception of a synthesis mode; the invention designs a key predictor as a guide to improve the generation precision of the atomic coordinates, and based on the strong dependence of the key on the atomic coordinates, the slight atomic disturbance can influence the type of the key and whether the key is formed, so that the cross entropy loss function is used and the gradient is used as an optimization direction to correct the generation of the atomic coordinates; the super-parameters for adding noise are predefined experience parameters, and can be optimized as an optimization target, so that the performance of the variable diffusion model such as E (3) in the generation of the inorganic crystal structure is comprehensively improved. Therefore, the invention not only has the innovation of the advantages, but also provides the optimization direction for other related researchers.
It should be noted that the method of the embodiments of the present disclosure may be performed by a single device, such as a computer or a server. The method of the embodiment can also be applied to a distributed scene, and is completed by mutually matching a plurality of devices. In the case of such a distributed scenario, one of the devices may perform only one or more steps of the methods of embodiments of the present disclosure, the devices interacting with each other to accomplish the methods.
It should be noted that the foregoing describes some embodiments of the present disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments described above and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
Example 2
Referring to fig. 2, embodiment 2 of the present invention further provides a crystal structure generating apparatus based on a diffusion model, including:
The data set acquisition module 1 is used for acquiring an open source crystal structure data set as a crystal structure generation data set, wherein the crystal structure generation data set is composed of cell parameters and CIF type files of atomic three-dimensional coordinates representing each crystal structure;
a format conversion module 2, configured to convert the CIF type file batch into an SDF type file required for training by using a chemical information file type conversion package, where the SDF type file includes a unit cell atom number, an atom bonding adjacency matrix, and a three-dimensional coordinate of each atom;
the data set dividing module 3 is used for dividing the crystal structure generated data set after format conversion into a training set, a verification set and a test set;
the diffusion model processing module 4 is used for constructing a crystal structure generation diffusion model, the crystal structure generation diffusion model is provided with a forward noise adding process and a reverse noise removing process, the forward noise adding process uses a predefined hyper-parameter to represent the mean value and variance of conditional probability distribution, and the reverse noise removing process restores data by predicting noise added by the forward noise adding process;
a key predictor construction module 5, configured to construct a key predictor through the graph neural network, where an input value of the key predictor is an atomic coordinate and an atomic type in the data set generated for the crystal structure, and an output value of the key predictor is a key type;
The model training module 6 is used for training the crystal structure generation diffusion model by using the training set, evaluating the prediction capability of the training crystal structure generation diffusion model in the unknown data set by using the verification set, and testing the trained crystal structure generation diffusion model by using the test set to obtain a final crystal structure generation diffusion model;
a crystal structure generation module 7, configured to generate a crystal structure by using a given gaussian noise and generating a diffusion model by using the constructed bond predictor and the final crystal structure.
And the linear interpolation processing module 8 is used for carrying out linear interpolation on the novel crystal structure generated by the diffusion model generated by the trained crystal structure, adding noise to different time stamps from two different crystal structures, and setting interpolation factors so as to control the occupation proportion of each crystal structure in the composite crystal.
In this embodiment, in the diffusion model processing module 4, diffusion processing is performed on the crystal structure generation data set after format conversion through the crystal structure generation diffusion model, and the crystal structure generation data set after diffusion processing approximately conforms to standard gaussian distribution;
in the process of inverse denoising of the diffusion model processing module 4, a neural network is constructed to learn and predict added noise, the relative distance between atoms is used for carrying out denaturation such as E (3) of a transmission characterization crystal, and an L2 norm between the added noise and the predicted noise is used as a loss function.
In this embodiment, in the key predictor construction module 5, the output value key type of the key predictor is determined as a classification task, the cross entropy is used as a loss function, the gradient value is used to guide the crystal structure to generate a learning conditional probability distribution of the diffusion model, and the generated crystal atomic coordinates are corrected;
in this embodiment, in the crystal structure generating module 6, noise values are randomly collected from the multidimensional standard gaussian noise distribution as a starting value to implement a reverse denoising process, and the gaussian noise dimension of the reverse denoising includes atomic coordinates contained in a unit cell and one-hot codes of all atomic types;
in the crystal structure generation module 7, the diffusion module is generated through the crystal structure to continuously denoise the input value according to the learned condition distribution, and finally, the true value conforming to the original data set distribution is obtained;
in the linear interpolation processing module 8, the crystal structure generation diffusion model is added through linear interpolation to generate crystal structure types, two different crystal structures are respectively noisy to time stamps T1 and T2, and t1+t2=t, wherein T is the total diffusion time step.
In this embodiment, the method further includes a model retraining module 9, configured to supplement the generated crystal structure to a training set, and retrain the diffusion model generated by the crystal structure to make the generated crystal structure conform to the distribution of the original data set.
It should be noted that, because the content of information interaction and execution process between the modules of the above-mentioned apparatus is based on the same concept as the method embodiment in embodiment 1 of the present application, the technical effects brought by the content are the same as the method embodiment of the present application, and specific content can be referred to the description in the foregoing illustrated method embodiment of the present application, which is not repeated herein.
Example 3
Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium having stored therein program code of a diffusion model-based crystal structure generation method, the program code including instructions for performing the diffusion model-based crystal structure generation method of embodiment 1 or any possible implementation thereof.
Computer readable storage media can be any available media that can be accessed by a computer or data storage devices, such as servers, data centers, etc., that contain an integration of one or more available media. The usable medium may be a magnetic medium (e.g., a floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid State Disk, SSD), etc.
Example 4
Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;
The processor and the memory complete communication with each other through a bus; the memory stores program instructions executable by the processor to invoke the program instructions capable of performing the diffusion model-based crystal structure generation method of embodiment 1 or any possible implementation thereof.
Specifically, the processor may be implemented by hardware or software, and when implemented by hardware, the processor may be a logic circuit, an integrated circuit, or the like; when implemented in software, the processor may be a general-purpose processor, implemented by reading software code stored in a memory, which may be integrated in the processor, or may reside outside the processor, and which may reside separately.
In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, produces a flow or function in accordance with embodiments of the present invention, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another, for example, by wired (e.g., coaxial cable, optical fiber, digital Subscriber Line (DSL)), or wireless (e.g., infrared, wireless, microwave, etc.).
It will be appreciated by those skilled in the art that the modules or steps of the invention described above may be implemented in a general purpose computing device, they may be concentrated on a single computing device, or distributed across a network of computing devices, they may alternatively be implemented in program code executable by computing devices, so that they may be stored in a memory device for execution by computing devices, and in some cases, the steps shown or described may be performed in a different order than that shown or described, or they may be separately fabricated into individual integrated circuit modules, or multiple modules or steps within them may be fabricated into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.
While the invention has been described in detail in the foregoing general description and specific examples, it will be apparent to those skilled in the art that modifications and improvements can be made thereto. Accordingly, such modifications or improvements may be made without departing from the spirit of the invention and are intended to be within the scope of the invention as claimed.
Claims (10)
1. The method for generating the crystal structure based on the diffusion model is characterized by comprising the following steps of:
acquiring an open source crystal structure data set as a crystal structure generation data set, wherein the crystal structure generation data set is composed of CIF type files representing unit cell parameters and atomic three-dimensional coordinates of each crystal structure;
converting the CIF type file batch into SDF type files required by training by using a chemical information file type conversion package, wherein the SDF type files comprise the number of unit cell atoms, an atom bonding adjacent matrix and three-dimensional coordinates of each atom; dividing the crystal structure generated data set after format conversion into a training set, a verification set and a test set;
constructing a crystal structure generation diffusion model, wherein the crystal structure generation diffusion model is provided with a forward noise adding process and a reverse noise removing process, the forward noise adding process uses predefined hyper-parameters to represent the mean value and variance of conditional probability distribution, and the reverse noise removing process restores data by predicting noise added by the forward noise adding process;
constructing a key predictor through a graph neural network, wherein the input value of the key predictor is the atomic coordinate and the atomic type in the data set generated by the crystal structure, and the output value of the key predictor is the key type;
Training the crystal structure generation diffusion model by using a training set, evaluating the prediction capability of the training crystal structure generation diffusion model in an unknown data set by using a verification set, and testing the trained crystal structure generation diffusion model by using a test set to obtain a final crystal structure generation diffusion model;
generating a diffusion model to generate a crystal structure by using a given Gaussian noise through the constructed bond predictor and the final crystal structure;
and performing linear interpolation on the novel crystal structure generated by the diffusion model generated by the trained crystal structure, adding noise to different time stamps from two different crystal structures, and setting interpolation factors to control the occupation proportion of each crystal structure in the composite crystal.
2. The diffusion model-based crystal structure generation method according to claim 1, wherein the crystal structure generation data set after format conversion is subjected to diffusion processing by the crystal structure generation diffusion model, and the crystal structure generation data set after diffusion processing conforms to standard gaussian distribution;
in the reverse denoising process, a neural network is constructed to learn and predict added noise, the relative distance between atoms is used for carrying out denaturation such as E (3) of a transmission characterization crystal, and the L2 norm between the added noise and the predicted noise is used as a loss function.
3. The diffusion model-based crystal structure generation method according to claim 1, wherein the output value of the bond predictor is determined as a classification task, the crystal structure is guided by gradient values using cross entropy as a loss function to generate a learning conditional probability distribution of the diffusion model, and the generated crystal atomic coordinates are corrected.
4. The method for generating a crystal structure based on a diffusion model according to claim 1, wherein, in the process of generating the crystal structure by generating the diffusion model from the final crystal structure by using a given gaussian noise:
randomly collecting noise values from multidimensional standard Gaussian noise distribution as a starting value to realize a reverse denoising process, wherein the Gaussian noise dimension of the reverse denoising comprises atomic coordinates contained in a unit cell and one-hot codes of all atomic types;
and continuously denoising the input value by the crystal structure generation diffusion module according to the learned conditional distribution, and finally obtaining a true value conforming to the distribution of the original data set.
5. The diffusion model-based crystal structure generation method according to claim 1, wherein the crystal structure generation diffusion model is increased by linear interpolation to generate a crystal structure type, two different crystal structures are respectively noisy to time stamps T1, T2, and t1+t2=t, T being a total diffusion time step.
6. The diffusion model-based crystal structure generation method according to claim 5, wherein the generated crystal structure is used as a training set to supplement, and the diffusion model generated by the crystal structure is trained again so that the generated crystal structure conforms to the distribution of the original data set.
7. A crystal structure generation device based on a diffusion model, comprising:
the system comprises a data set acquisition module, a crystal structure generation module and a crystal structure generation module, wherein the data set acquisition module is used for acquiring an open source crystal structure data set as a crystal structure generation data set, and the crystal structure generation data set is composed of cell parameters representing each crystal structure and CIF type files of atomic three-dimensional coordinates;
the format conversion module is used for converting the CIF type file batch into SDF type files required by training by using a chemical information file type conversion package, wherein the SDF type files comprise the number of unit cell atoms, an atom bonding adjacent matrix and three-dimensional coordinates of each atom;
the data set dividing module is used for dividing the crystal structure generated data set after format conversion into a training set, a verification set and a test set;
the diffusion model processing module is used for constructing a crystal structure generation diffusion model, the crystal structure generation diffusion model is provided with a forward noise adding process and a reverse noise removing process, the forward noise adding process uses a predefined hyper-parameter to represent the mean value and variance of conditional probability distribution, and the reverse noise removing process restores data by predicting noise added by the forward noise adding process;
The key predictor construction module is used for constructing a key predictor through the graph neural network, wherein the input value of the key predictor is the atomic coordinate and the atomic type in the data set generated for the crystal structure, and the output value of the key predictor is the key type;
the model training module is used for training the crystal structure generation diffusion model by using the training set, evaluating the prediction capability of the training crystal structure generation diffusion model in an unknown data set by using the verification set, and testing the trained crystal structure generation diffusion model by using the test set to obtain a final crystal structure generation diffusion model;
a crystal structure generation module for generating a crystal structure by using a given gaussian noise through a constructed bond predictor and the final crystal structure generation diffusion model;
and the linear interpolation processing module is used for carrying out linear interpolation on the novel crystal structure generated by the diffusion model generated by the trained crystal structure, adding noise to different time stamps from two different crystal structures, and setting interpolation factors so as to control the occupation proportion of each crystal structure in the composite crystal.
8. The diffusion model-based crystal structure generation apparatus according to claim 7, wherein in the diffusion model processing module, diffusion processing is performed on the crystal structure generation data set after format conversion by the crystal structure generation diffusion model, and the crystal structure generation data set after diffusion processing conforms to a standard gaussian distribution;
In the reverse denoising process of the diffusion model processing module, a neural network is constructed to learn and predict added noise, the relative distance between atoms is used for carrying out denaturation such as E (3) of a transmission characterization crystal, and the L2 norm between the added noise and the predicted noise is used as a loss function.
9. The diffusion model-based crystal structure generation apparatus according to claim 7, wherein the key predictor construction module determines the type of the key output value of the key predictor as a classification task, guides the crystal structure through a gradient value using cross entropy as a loss function to generate a learning condition probability distribution of the diffusion model, and corrects the generated crystal atomic coordinates.
10. The crystal structure generating device based on the diffusion model according to claim 7, wherein in the crystal structure generating module, noise values are randomly collected from a multi-dimensional standard Gaussian noise distribution as initial values to realize a reverse denoising process, and the Gaussian noise dimension of the reverse denoising comprises atomic coordinates contained in a unit cell and one-hot codes of all atomic types;
in the crystal structure generation module, the diffusion module is generated through the crystal structure to continuously denoise the input value according to the learned condition distribution, and finally, the true value conforming to the distribution of the original data set is obtained;
In the linear interpolation processing module, the crystal structure generation diffusion model is added through linear interpolation to generate crystal structure types, two different crystal structures are respectively noisy to time stamps T1 and T2, and t1+t2=t, wherein T is the total diffusion time step;
the model retraining module is used for supplementing the generated crystal structure into a training set, and retraining the diffusion model generated by the crystal structure so that the generated crystal structure accords with the distribution of the original data set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311737478.8A CN117423396B (en) | 2023-12-18 | 2023-12-18 | Crystal structure generation method and device based on diffusion model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311737478.8A CN117423396B (en) | 2023-12-18 | 2023-12-18 | Crystal structure generation method and device based on diffusion model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117423396A true CN117423396A (en) | 2024-01-19 |
| CN117423396B CN117423396B (en) | 2024-03-08 |
Family
ID=89528744
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311737478.8A Active CN117423396B (en) | 2023-12-18 | 2023-12-18 | Crystal structure generation method and device based on diffusion model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117423396B (en) |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060173633A1 (en) * | 2005-01-28 | 2006-08-03 | Ceptyr, Inc. | Crystalline phosphatase and method for use thereof |
| CN114141313A (en) * | 2021-10-28 | 2022-03-04 | 西安理工大学 | Numerical simulation method for predicting liquid/liquid two-phase interface reaction phase change process |
| CN114334028A (en) * | 2022-01-05 | 2022-04-12 | 哈尔滨工业大学（深圳） | Method, system and storage medium for generating indefinite 3D structure based on graph neural network |
| CN115206457A (en) * | 2022-07-22 | 2022-10-18 | 腾讯科技（深圳）有限公司 | Three-dimensional molecular structure generation method, device, equipment and storage medium |
| CN115394365A (en) * | 2022-08-31 | 2022-11-25 | 深圳晶泰科技有限公司 | Crystal structure similarity calculation method, screening method, device and equipment |
| CN116130018A (en) * | 2023-01-04 | 2023-05-16 | 澳门大学 | Organic crystal structure prediction method, device, equipment and storage medium |
| CN116304705A (en) * | 2023-03-21 | 2023-06-23 | 南京邮电大学 | Flow data set generation method and device based on conditional diffusion model |
| CN116343937A (en) * | 2022-12-28 | 2023-06-27 | 深圳晶泰科技有限公司 | Crystal structure generation method, crystal structure generation device, electronic equipment and computer readable storage medium |
| CN116913414A (en) * | 2023-07-14 | 2023-10-20 | 武汉大学 | Fusion lipophilic drug molecule generation method and system based on molecule generation model |
| CN116994642A (en) * | 2023-08-09 | 2023-11-03 | 上海交通大学 | Discrete graph probability denoising diffusion model for protein sequence generation |
| CN117198417A (en) * | 2023-10-17 | 2023-12-08 | 西安工业大学 | Stable crystal structure prediction method and system based on machine learning and target optimization |
-
2023
- 2023-12-18 CN CN202311737478.8A patent/CN117423396B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060173633A1 (en) * | 2005-01-28 | 2006-08-03 | Ceptyr, Inc. | Crystalline phosphatase and method for use thereof |
| CN114141313A (en) * | 2021-10-28 | 2022-03-04 | 西安理工大学 | Numerical simulation method for predicting liquid/liquid two-phase interface reaction phase change process |
| CN114334028A (en) * | 2022-01-05 | 2022-04-12 | 哈尔滨工业大学（深圳） | Method, system and storage medium for generating indefinite 3D structure based on graph neural network |
| CN115206457A (en) * | 2022-07-22 | 2022-10-18 | 腾讯科技（深圳）有限公司 | Three-dimensional molecular structure generation method, device, equipment and storage medium |
| CN115394365A (en) * | 2022-08-31 | 2022-11-25 | 深圳晶泰科技有限公司 | Crystal structure similarity calculation method, screening method, device and equipment |
| CN116343937A (en) * | 2022-12-28 | 2023-06-27 | 深圳晶泰科技有限公司 | Crystal structure generation method, crystal structure generation device, electronic equipment and computer readable storage medium |
| CN116130018A (en) * | 2023-01-04 | 2023-05-16 | 澳门大学 | Organic crystal structure prediction method, device, equipment and storage medium |
| CN116304705A (en) * | 2023-03-21 | 2023-06-23 | 南京邮电大学 | Flow data set generation method and device based on conditional diffusion model |
| CN116913414A (en) * | 2023-07-14 | 2023-10-20 | 武汉大学 | Fusion lipophilic drug molecule generation method and system based on molecule generation model |
| CN116994642A (en) * | 2023-08-09 | 2023-11-03 | 上海交通大学 | Discrete graph probability denoising diffusion model for protein sequence generation |
| CN117198417A (en) * | 2023-10-17 | 2023-12-08 | 西安工业大学 | Stable crystal structure prediction method and system based on machine learning and target optimization |
Non-Patent Citations (2)
| Title |
|---|
| V. VASHOOK 等: "Crystal Structure, Oxygen Nonstoichiometry and Diffusion Mobility in Some Ferrites-Nickelites", 《OMEE-2012》, 31 December 2012 (2012-12-31), pages 105 - 106 * |
| 薛少宗 等: "不同晶体形貌SAPO-34分子筛合成方法研究进展", 《应用化工》, vol. 47, no. 8, 31 August 2018 (2018-08-31), pages 1741 - 1748 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117423396B (en) | 2024-03-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Han et al. | Unsupervised generative modeling using matrix product states | |
| US10909418B2 (en) | Neural network method and apparatus | |
| Li et al. | AI-assisted superresolution cosmological simulations | |
| Li et al. | Neural network renormalization group | |
| Chen et al. | Big data deep learning: challenges and perspectives | |
| Nguyen et al. | Quality not quantity: On the interaction between dataset design and robustness of clip | |
| US20200302271A1 (en) | Quantization-aware neural architecture search | |
| Tang et al. | Implementation of deep learning-based automatic modulation classifier on FPGA SDR platform | |
| CN113434683B (en) | Text classification method, device, medium and electronic equipment | |
| Afolabi et al. | Hierarchical meta-learning in time series forecasting for improved interference-less machine learning | |
| Hrúz et al. | One model is not enough: Ensembles for isolated sign language recognition | |
| Didi et al. | Wavelet density and regression estimators for continuous time functional stationary and ergodic processes | |
| CN117423396B (en) | Crystal structure generation method and device based on diffusion model | |
| Jiang et al. | Modulation recognition of communication signal based on convolutional neural network | |
| Khan et al. | Efficiently processing big data in real-time employing deep learning algorithms | |
| Sheng et al. | LA-ESN: a novel method for time series classification | |
| Stoican et al. | Aiding dictionary learning through multi-parametric sparse representation | |
| CN114281950B (en) | Data retrieval method and system based on multi-graph weighted fusion | |
| Duan et al. | An efficient algorithm for solving the nonnegative tensor least squares problem | |
| Mushtaq et al. | Super Resolution for Noisy Images Using Convolutional Neural Networks | |
| Wei et al. | A family of automatic modulation classification models based on domain knowledge for various platforms | |
| Sun et al. | Simultaneous smoothing and untangling of 2D meshes based on explicit element geometric transformation and element stitching | |
| Nauman et al. | Exploring helical dynamos with machine learning: Regularized linear regression outperforms ensemble methods | |
| Yimamu et al. | Robust estimation of highly corrupted low‐rank matrix via alternating direction method of multiplier | |
| Liu et al. | Radio Frequency Database Construction and Modulation Recognition in Wireless Sensor Networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |